On a fixed cycle measured in months, not years, and straight away after any change to the agent's model, sources, brief or tools. Between those points, scheduled test runs and live monitoring should be able to pull a certificate early. Higher depth, faster-changing sources and higher volume all argue for a shorter cycle.
Why months rather than years?
Because the things that drive an agent's behaviour change on that timescale. Model providers release updates. Policy documents are revised. Products launch and are withdrawn. Customers start asking new kinds of questions. Each one can move the agent away from the behaviour it was certified on, often without anyone noticing.
A yearly certificate would spend most of its life describing an agent that no longer exists. A cycle of a few months keeps evidence close to reality while giving teams time to do the work properly.
What is the difference between retesting and recertifying?
They are related but not the same.
| Retesting | Recertifying | |
|---|---|---|
| What it is | Running the agent on its test set | Issuing a fresh certificate on new evidence |
| How often | Frequently, often on a schedule between renewals | On the fixed cycle, or after a change |
| Who acts | Automated runs, checked by operations | Test team, with an independent reviewer |
| What a failure does | Drops the agent's depth the same day | Certificate is not reissued until it passes |
| Updates the case set? | Usually not | Yes, adds new cases from operation |
Retesting is the early-warning system; recertification is the formal renewal. Retesting cadence is covered in how often an AI agent should be retested.
What should shorten the cycle?
- Higher depth. An agent that acts on its own, especially one that moves money, needs fresher evidence than one that drafts for a person.
- Fast-changing sources. Agents reading pricing, product terms or regulatory guidance that changes often.
- High volume. More actions mean a drift costs more before it is caught.
- Customer-facing work. Errors reach people directly.
- A recent failure. After a failed run and a fix, a shorter next cycle confirms the fix holds.
What can lengthen it a little: a read-only agent, stable sources, low volume, and a long run of clean retests. Even then, a fixed outer limit stops certificates from going stale by default.
What should trigger recertification outside the cycle?
Any change to what produces the agent's behaviour:
- a new model or model version;
- a source added, removed or replaced with a new version;
- an edit to the brief or instructions;
- a new tool, or a change to a tool's limits;
- a request to raise a capability's depth.
Each starts a rerun. If it passes, the certificate is reissued and the cycle restarts.
What should recertification look at beyond the scores?
Patterns from operation. If the agent has been refusing or escalating far more than expected, that is a signal even if the test set passes; see how to handle an AI agent that refuses too often. If people reviewing held items have been overriding the agent often, the reasons belong in the next case set.
Reviewers should also check whether the sources themselves are still right. An agent can pass the grounded gate perfectly while repeating an out-of-date policy, which is the point made in grounded is not the same as correct.
An illustrative example
Take an asset manager with two agents. One answers internal staff questions about fund documentation at depth 1. The other drafts responses to client redemption requests at depth 2 and processes routine ones at depth 3.
The documentation agent runs on a quarterly cycle. Its sources change only when fund documents are reissued, which triggers a rerun anyway. The redemption agent also runs on a quarterly outer limit, but with weekly retests, because its sources include daily dealing notices. Midway through one quarter, a weekly retest clears no depth after a change to the dealing cut-off. The agent drops to read-only that day, the brief is updated, and it is recertified within the week. Its next full renewal comes due on the new date.
Who decides the cycle?
The process owner, with risk input, recorded on the certificate. Changing the cycle should be a recorded decision, not a quiet extension. The thinking behind certification in general is in what it means to certify an agent.
Where Pryme fits
In Pryme, every certificate lapses after 90 days, so the outer limit is fixed. Between renewals, an Attest run that falls short keeps only the highest depth it cleared, one that clears no depth drops the agent to depth 1 the same day, and any change to the model, sources or brief means a fresh run before the certificate counts again.
Questions people ask
Is recertification expensive?
Less than the first certification, because the case set and method already exist. Most of the cost is reviewing failures and adding new cases.
Can recertification be fully automated?
The runs can. The acceptance should stay with a named person who reads the results, including changes from the last run.
What if recertification is overdue?
The certificate lapses and the agent drops to read-only until it passes. An overdue certificate should not be treated as still valid.